MiniFAS Face Anti-spoofing System

Quantizing the model to INT8 provides significant deployment advantages with negligible impact on performance:

• Accuracy: No measurable drop in accuracy compared to the FP32 model.
• Size Reduction: The file size is reduced to approximately 600 KB, which is 67% smaller than the original.
• Performance Parity: Maintains the same ROC-AUC and Average Precision (AP) scores as the FP32 version.

The pipeline uses different interpolation methods depending on whether the face crop is being upscaled or downscaled to reach the target 128x128 resolution:

• Upscaling (Small crop $\rightarrow$ 128x128): Uses cv2.INTER_LANCZOS4. This high-quality interpolation uses a windowed sinc function to ensure sharp edges and preserve texture without blocky artifacts.
• Downscaling (Large crop $\rightarrow$ 128x128): Uses cv2.INTER_AREA. This method averages pixels that map to each output pixel, providing proper downsampling that prevents aliasing and preserves texture better than bilinear or bicubic interpolation.

interp = cv2.INTER_LANCZOS4 if crop_size < target_size else cv2.INTER_AREA

The model's accuracy is affected by the following environmental and physical factors:

• Pose: Accuracy is highest with frontal views (within +/-30 degrees of yaw/pitch). Profile views significantly reduce performance.
• Occlusions: Masks, hands covering the face, or thick glasses with heavy reflections interfere with texture extraction.
• Attack Types: The model is effective against printed photos and screens but is not trained on 3D silicone masks or prosthetics.
• Motion Blur: Fast movement smears textures. To mitigate this in video streams, implement temporal filtering by requiring a sequence of consecutive 'Real' predictions.

The Fourier Transform branch is an auxiliary component used only during training to improve the backbone's ability to detect frequency-based spoofing artifacts (like screen pixel grids or print halftoning).

Mechanism:

• It takes features from Stage 2.
• It uses three convolutional layers (128→64→1) to predict a single-channel Fourier spectrum.
• The prediction is supervised against the actual FFT of the input image.
• Gradients from this branch flow back into the main backbone.

Note: This branch is removed during inference, so it does not add overhead to the deployed model.

SE blocks are used at the boundaries of each stage to allow the network to dynamically emphasize specific feature channels that are most relevant to the input (e.g., channels detecting moiré patterns vs. those detecting paper texture).

The SE process:

1. Squeeze: Global average pooling is applied to obtain one value per channel.
2. Excitation: Two small Fully Connected (FC) layers process these values to produce a weight for each channel.
3. Scale: The original feature maps are multiplied by these calculated weights.

MiniFASNet V2 SE is a specialized neural network designed for face anti-spoofing (distinguishing real faces from fakes like printed photos, screens, or masks) rather than general image classification.

It achieves this by detecting subtle presentation attack artifacts such as:

• Moiré patterns from phone screens.
• Color shifts from printed photos.
• Texture differences between paper/plastic and skin.
• Reflections and glare from displays.

To maintain a small footprint while remaining effective, the architecture uses:

1. Depthwise separable convolutions to minimize parameter count.
2. Squeeze-and-Excitation (SE) blocks to prioritize relevant feature channels.
3. Fourier Transform (FT) supervision (during training) to force the network to learn frequency-domain patterns characteristic of spoofs.

To train the model, you need face images (.jpg or .png), bounding box files (image_BB.txt with x y w h), and label files (metas/labels/train_label.json and metas/labels/test_label.json).

Use scripts/prepare_data.py to crop and resize faces into organized training and testing folders:

python scripts/prepare_data.py \
  --orig_dir <path> \
  --crop_dir <path> \
  --size <number> \
  --bbox_expansion_factor <float> \
  --spoof_types <number> [<number> ...]

Run the training script using the prepared cropped dataset. Checkpoints and TensorBoard logs are saved to <output_dir>/MINIFAS/.

Standard Training:

python scripts/train.py \
  --crop_dir <path> \
  --input_size <number> \
  --batch_size <number> \
  --output_dir <path>

Resume Training:

python scripts/train.py \
  --crop_dir <path> \
  --resume <checkpoint_path>

To use MiniFAS, you must use Python 3.8.0 or higher. Note that Python 3.7.x is incompatible and will cause installation errors during dependency setup.

1. Create and activate a virtual environment

Using Conda:

conda create -n face-antispoof python
conda activate face-antispoof

Using venv:

python -m venv venv
# Linux/macOS
source venv/bin/activate
# Windows
venv\Scripts\activate

2. Install dependencies

pip install -r requirements.txt

IMPORTANT

To run inference on a GPU, install onnxruntime-gpu instead of onnxruntime.